Hacker News new | ask | show | jobs
by solveit 456 days ago
> I'd argue that the biggest reason machines are black boxes are because no one is bothering to look inside of them.

People do look, but it's extremely hard. Take a look at how hard the mechanistic interpretability people have to work for even small insights. Neel Nanda[1] has some very nice writeups if you haven't already seen them.

[1]: https://www.neelnanda.io/mechanistic-interpretability
2 comments
godelski 456 days ago 

  > People do look
This was never in question

  > Very few are trying to understand why things are working
What is in question is why this is given so little attention. You can hear Neel talk about this himself. It is the reason he is trying to rally people and get more into Mech Interp. Which frankly, this side of ML is as old as ML itself.

Personal, I believe that if you aren't trying to interpret results and ask the why then you're not actually doing science. Which is fine. There's plenty of good things that come from outside science. I just think it's weird to call something science if you aren't going to do hypothesis testing and finding out why things are the way they are

link
jebarker 456 days ago
The problem is that mechanistic interpretability is a lot like neuroscience or molecular biology, i.e. you're trying to understand huge complexity from relatively crude point measurements (no offense intended to neuroscientists and biologists). But AI wants publishable results yesterday. I often wonder whether the current AI systems will stay around long enough for anyone to remain interested in understanding why they ever worked.
link
godelski 456 days ago
People will always be interested in why things work. At least one will as long as I'm alive, but I really don't think I'm that special. Wondering why things are the way they are is really at the core of science. Sure, there are plenty of physicists, mathematicians, neuroscientists, biologists, and others who just want answers, but this is a very narrow part of science.

I would really encourage others to read works that go through the history of the topic they are studying. If you're interested in quantum mechanics, the one I'd recommend is "The Quantum Physicists" by William Cropper[0]. It won't replace Griffiths[1] but it is a good addition.

The reason that getting information like this is VERY helpful is that it teaches you how to solve problems and actually go into the unknown. It is easy to learn things from a book because someone is there telling you all the answers, but texts like these instead put yourself in the shoes of the people in those times, and focus on seeing what and why certain questions are being asked. This is the hard thing when you're at the "end". When you can't just read new knowledge from a book, because there is no one that knows! Or the issue Thomas Wolf describes here[2] and why he struggled.

[0] https://www.amazon.com/Quantum-Physicists-Introduction-Their...

[1] https://www.amazon.com/Introduction-Quantum-Mechanics-David-...

[2] https://thomwolf.io/blog/scientific-ai.html

link